Where Model-Driven Screen Interaction Actually Runs Out of Room

Community Article
Published August 18, 2026

A question this community will appreciate more than most: when a model-driven agent operates a graphical interface, what determines where its usefulness ends — the model, or the surface it's allowed to act on?

Browser agents like Comet (Perplexity) are a good case study. They're capable — genuinely strong at page-level Q&A, tab-context synthesis, and supported web actions — but their control surface is bounded to the browser session: DOM, tabs, browser-accessible documents. No amount of model capability changes that boundary; it's architectural, not a model-quality problem.

The observation/action loop, one layer down

Aiden approaches this from the device layer instead of the browser layer. The documented dev-board implementation captures a connected screen via HDMI, and sends input via USB HID — meaning the "interface" being modeled isn't DOM structure or accessibility labels, it's raw pixels, same as any other real-world vision task.

That has real implications for the perception problem:

  • No DOM to fall back on for grounding — button detection, layout interpretation, and text extraction all have to work from the same visual signal a human would use
  • State tracking matters as much as single-frame interpretation — a button correctly identified in one frame can still fail if a dialog appeared, a screen is mid-load, or an earlier action silently didn't take effect
  • The interaction loop is deliberately conservative: observe → interpret → propose one bounded action → verify the resulting state → escalate to the user when ambiguous, rather than chaining actions on unverified assumptions
flowchart LR
    A[User instruction] --> B{Current control surface}
    B --> C[Browser session]
    C --> D["Tabs, webpages, web apps"]
    B --> G[Connected device interface]
    G --> H["Browser, native app, desktop app, or system dialog"]
    H --> I[Observe visible state]
    I --> J[Propose next action]
    J --> K[User confirms, redirects, or stops]
    K --> L[Execute supported interaction]
    L --> M[Review outcome]

Why this is a genuinely different category, not a Comet knockoff

Worth being precise here: this isn't "Aiden vs. Comet." Fellou is explicitly agentic ("self-driving browser"), Dia has strong cross-tab context modeling (though more suggestive than autonomous in practice), Opera Neon is doing something more experimental around creation workflows. All of these are still operating within a browser's DOM/tab model. Aiden's distinguishing architectural choice is dropping down to raw screen capture + HID input specifically so the target interface doesn't need to be a browser at all — a native app, a desktop tool, a system dialog, anything with a visible screen and an input path, is fair game architecturally, even though real support today is bounded to documented Android/iPhone workflows.

The part relevant to grounding/vision folks specifically

The bring-your-own-multimodal-model design means the perception quality is entirely a function of whatever model you point it at — this is deliberately not a fine-tuned, task-specific vision model. That's a real tradeoff: you get to swap in whatever's strongest for your use case, but you also inherit whatever grounding weaknesses that general-purpose model has on UI-specific tasks (small text, overlapping elements, near-identical icons, mid-animation frames).

Curious whether others here have compared general-purpose multimodal models against UI-specialized grounding models for this kind of task, and at what latency/accuracy tradeoff.

Firmware (public): github.com/AidenAI-IO/aiden-firmware

Discord for technical discussion: discord.com/invite/bcJavjcnYz

Community

Sign up or log in to comment